Fabric defect image synthesis using GANs: quality assessment and defect detection performance
Karthika Latha, Arulanand Natarajan, Viju SubramoniapillaiPurpose
The purpose of this study is to use generative adversarial networks (GANs) to produce high-quality synthetic fabric defect images and evaluate their quality using the Fréchet Inception Distance (FID) score. This study also aims to assess fabric defects synthesized using GANs with an EfficientNet-based model.
Design/methodology/approach
In this study, three types of GAN frameworks, namely, deep convolutional GAN (DCGAN), Wasserstein GAN (WGAN) and dual generator GAN (DG2GAN), were used to produce synthetic images of fabric defects, including holes, stains, metal contamination and thread errors. The quality of the generated images was measured using an FID metric. Furthermore, the generated fabric defect images were evaluated using an EfficientNet classifier to study their effectiveness for defect identification. Gradient-weighted class activation mapping (Grad-CAM) visualizations were used to assess the model’s attention toward relevant defect regions.
Findings
The results showed that DG2GAN generated the most realistic synthetic defect images with the lowest FID scores, ranging from 12 to 38 and was closer to the real defect samples. In comparison, DCGAN and WGAN produced higher FID scores, indicating lower fidelity of the images and less stable generation performance. The EfficientNet classifier achieved a high classification accuracy of 93.5%, and feature visualization using Grad-CAM revealed that the classifier focused on defect-specific visual features.
Originality/value
In this study, high-quality images of fabric defects were generated using the DG2GAN architecture, which demonstrated enhanced image realism and achieved a notably low FID score. The generated synthetic images were used to train an EfficientNet-based classifier, which exhibited a high detection accuracy and validated the effectiveness of the synthetic data generated for fabric defect detection.